Pinterest uses Pinnability, a machine learning system, to rank every Pin shown on each user's home feed based on predicted engagement. The system has been in place since 2015, replacing a chronological ordering of content, and is continuously updated; the most recent major version, TransActV2, was published as a production model in June 2025 and extends personalization to up to 16,000 historical user actions per person.

The system operates in two stages. First, a retrieval layer uses over 20 candidate generators — including embedding-based nearest-neighbor search across billions of Pins — to assemble a pool of potentially relevant content for each user. Then a ranking layer uses deep neural networks to score each candidate using combinations of Pin signals, long-term user preference embeddings (mathematical representations of a user's tastes built from their entire save history), and real-time signals reflecting what the user has done in the current session. TransActV2 extended the historical action window from 100 to 16,000 actions per user, running on custom GPU kernels to maintain response times under 100 milliseconds.